0820 | Origin by Cursor, Hosted Agents in Cluing, ChatGPT for Teens, AgentR 3.0

Mia: Welcome back to ProductHunt Daily on Bri Radio. I'm Mia, and alongside me, Milo. Today we're tracking what's been launching across the Product Hunt graph — from a repairable phone built to outlast this entire decade, to an AI hiring agent that actually conducts interviews, plus a dedicated ChatGPT experience built for teenagers. Let's get into it.

Milo: We've also got a self-hosted, privacy-first analytics tool that runs on a single Cloudflare Worker, a text-to-speech studio pitched as an ElevenLabs alternative, and an AI workspace for reproducible scientific research coming out of Peru. Milan, where do you want to start?

Mia: Cursor has launched Origin, a new code hosting platform it calls a Git forge built for the age of coding agents. Origin lets you create and host Git repositories right inside Cursor, browse and search code, open and review pull requests, manage access, and sync with GitHub. Cursor frames it for a world where humans and coding agents work side by side, so your code lives next to the agents working on it, with more agent-native features still on the way.

Mia: Two new products are pushing agents out of a single personal device and into a shared workspace. Cluing's Hosted Agents let you start work on your computer, continue on your phone, and hand it to a teammate without losing context, with skills that update as you collect new insights. Anyone can assign tasks, publish work, and connect to external tools through an API or MCP using Cluing's custom connectors.

Milo: That evolving skill really changes what an agent is. It stops being a tool you stay tethered to and becomes a team asset that keeps working once you close the laptop. OmniVibe chases the other side of that same shift, a marketplace where creators and users come together. Creators can import and publish agents they already have and earn from qualified usage, while users can browse or describe what they need and work with multiple agents at once.

Mia: And that creator economy framing is the exact counterpart to Cluing. One product turns agents into a persistent team asset, the other gives those agents a home beyond a local harness, a GitHub repo, or a skills file, and a way for their makers to get paid.

Mia: OpenAI has introduced ChatGPT for Teens, a dedicated version for ages 13 to 17, designed around how teenagers learn and use AI, according to the product listing. It's meant to encourage working through problems instead of just handing over a homework answer, and it includes Study Mode, quizzes, learning visualizations, and Study Hours. The listing claims stronger safety protections and healthy-use features are built in by default, with optional parental controls, while keeping teens' conversations private.

Milo: But those capabilities are all maker descriptions straight from the product page, so the concrete behavior here is untested. No independent verification or third-party test results were provided, which means how well Study Mode and the safety protections actually perform in a teenager's real day-to-day use remains unverified. On top of that, OpenAI hasn't disclosed pricing or availability yet.

Mia: AgentR is on version 3.0, and the company lays out a compact division of labor: it reads, it checks, it interviews, and you still decide. Every application gets read, candidate claims get verified, then an AI runs an interview, aimed at hiring teams running high-volume hiring, where the maker says the problem shows up first. The maker frames the release around a shifting problem, losing trust in virtual interviews. The first version targeted noisy resumes, version two targeted disconnected hiring workflows, and version three targets candidates using AI at every stage of an interview, which the maker says most hiring tools have no way to detect.

Milo: So what's new in this phase centers on that first interview. It's a structured 25-minute voice-and-video conversation that adapts to what the candidate says rather than following a fixed script, covering a resume walkthrough, real case scenarios, judgment calls, and hands-on problem solving. Verification comes before evaluation, with a deep research engine checking resumes for consistency before they're scored.

Mia: And the anti-cheat approach is designed rather than monitored, with the product page describing two proctoring layers, one watching the room and one watching the machine. Any flag lands as a marker on the recording, never as a score. The maker says rehearsed answers and real-time AI help don't hold up, and candidates are then ranked on how their experience fits the role. The maker explicitly asks for feedback from teams running high-volume hiring.

Milo: There's a new tool aimed at fixing one of the messiest parts of doing science. Ressearch AI, built by Skyfall Innovations in Peru, is an AI workspace for reproducible research. Founder Irwing says the product grew out of researchers losing momentum as they bounce between papers, search engines, their Python or R code, statistical tools, and Word — every handoff dropping context and making the final work harder to review or reproduce.

Mia: Mm. So it's one conversational workspace covering literature search, data acquisition, the analysis, visualizations, editing, and scientific writing. The AI agents run traceable workflows in isolated cloud sandboxes, and the whole thing preserves sources, assumptions, code, decisions, and results, then keeps everything exportable to GitHub or locally with no setup required.

Milo: Irwing's headline claim is that specialized agents can search literature, run the Python or R analyses, create traceable figures, tables, maps, and even 3D scientific objects, and then draft and review the writing grounded in the project evidence. Though he's careful to say the goal is keeping the researcher in control, not replacing scientific judgment.

Mia: There's actually a discrepancy in the available material on how many sources it connects to. Irwing said in the launch comment that it reaches more than thirty scientific sources, but the product website says sixty, naming the likes of Semantic Scholar, OpenAlex, PubMed, arXiv, Crossref, and ClinicalTrials.gov. Either way, it's aimed at researchers, graduate students, labs, and scientific teams across the life sciences, from genomics and microbiology to drug discovery and ecology.

Milo: Fairphone has a new model out, and it's a deliberately incremental one. The Gen 6 Plus is a repairable 5G Android phone the maker says is built to last through 2033, with a five-year warranty, replaceable parts, and a 50-megapixel camera.

Mia: Community discussion adds the key detail: it keeps the same repairable chassis as last year's Fairphone 6, with twelve user-replaceable parts and a removable battery. The real update is internal — a Snapdragon 7s Gen 4 chipset and 12GB of RAM, which one commenter said gives a phone meant to last for years more headroom. They also described the Plus as deliberately modest, with no bigger display or extra camera than the standard 6.

Milo: It actually ships with a screwdriver in the box, and that same commenter stressed that buyers are genuinely expected to open the phone. For the first time, Fairphone is also selling directly in the US. People on the launch thread picked up on real specifics — one said it's rare to see privacy-related products on Product Hunt, and another said the appeal is swapping in a fresh battery instead of living with a dying iPhone one.

Mia: But there's a genuine open question in the launch discussion about whether repairability will land with American buyers the way it has in Europe. The supplied launch material gives no data on US reception yet, so that stays untested.

Milo: Two hardware tools debuted that sit at different points in the electronics workflow. Cherry Blossom is a browser app that builds fab-ready custom PCBs from a single plain-English prompt — it generates the schematic, the PCB layout, the bill of materials, and the fabrication files in about five minutes at what it calls electrical-engineering grade. For launch day they're giving away a million tokens free.

Mia: And the community thread is already piling on a competitor — one comment says they'd had enough of flux.ai's product after it stopped working and cost them fifty dollars. That's the space Cherry Blossom is trying to undercut.

Milo: KiHub, meanwhile, is a hardware review platform built for KiCad projects specifically. It lets teams compare schematic and PCB revisions, discuss changes with pinned comments, run automated checks, manage approvals, and keep a traceable review history — all without changing their existing Git workflow. So Cherry Blossom handles the build, while KiHub covers the review side of the same pipeline.

Milo: Finally, Vois 2.0 is a desktop text-to-speech studio pitched as an ElevenLabs alternative, built by solo developer Praney in Melbourne. His stated problem is what he calls the preview tax: cloud voice tools charge per character, so every retake of the same paragraph costs money again and quietly discourages iteration. Vois runs locally and paid subscriptions include unlimited generation — no tokens, no usage meter, no per-character fees.

Mia: The release includes four engines. Fast is aimed at drafts at around three times real time on CPU and six times with GPU acceleration on Apple Silicon; then there's Expressive, a multilingual engine covering 23 languages, and a Pro-only Omni offering 600-plus languages with voice design. It also brings 100-plus voices across 21 categories, voice cloning from about fifteen seconds of clean audio with a consent step, a multi-speaker timeline with music and mastering, and a CLI that agents like Claude, Codex, or Cursor can drive.

Milo: Since version 2.0 it's added encrypted portable backups and reconstructible generation records, and the privacy pitch holds because scripts, voice samples, and generated audio all stay on the user's machine. Launch pricing is ten dollars a month for the Subscriber tier and fourteen for Pro for the first hundred spots, locked in for the life of the subscription and ending August 31st, with a seven-day free trial that needs no card.

Milo: Edgemetry positions itself as the middle ground for site analytics: a self-hosted, privacy-first tool that runs on a single Cloudflare Worker and one D1 database on Cloudflare's free tier, with no cookies and no consent banner. The maker built it for sites where Google Analytics feels like overkill and drags a consent banner along, Plausible Cloud is a recurring bill, and self-hosting Plausible means running a VPS with Docker — more infrastructure than a personal site deserves.

Mia: Deployment really is one click. Cloudflare signs you in, copies the repository into your own GitHub or GitLab account, creates the D1 database automatically, and deploys the Worker — no API token to generate. You point a subdomain at it and add a roughly 2.1 kilobyte tracking script served from your own domain. Just know a freshly created Worker has no route enabled until you give it a hostname.

Milo: Tracking covers visitors, pageviews, views per visit, bounce rate and average time, each with a sparkline and a previous-period comparison. Filters are stackable — clicking a row narrows every panel at once — and breakdowns cover page, entry and exit page, referrer, country, browser, OS, device, screen size and UTM tags. One deployment supports multiple sites and users with per-site access.

Mia: The real care seems to have gone into details. The map is a choropleth projected and served from the Worker itself, so there's no CDN call. The realtime view shows a live counter, the last thirty minutes minute by minute, and the pages being read right now. Custom events use a one-line call, and a command palette searches every dimension and returns filters rather than links. Daily rollups are never deleted, so history is unlimited — and it all stays inside your own Cloudflare account rather than handing data to a third party.

Mia: Claude Watermark Remover is a browser-based tool that finds and removes the traces a chat interface leaves in text. You paste text and it identifies hidden HTML class names, zero-width characters, exotic spaces, and typography — each finding with a count and a position, because these are facts about the bytes rather than a probability score. Then you clean it in one click.

Milo: It draws a deliberate boundary in what it claims to do. It does not try to detect Anthropic's statistical watermark, because the creator says nobody outside Anthropic can, so the tool restricts itself to the structural traces chat interfaces leave behind.

Mia: It's free, unlimited, needs no account, and runs entirely in your browser so nothing gets uploaded. The engine is MIT open source, which means anyone can inspect exactly how it works.

Mia: So we've covered a lot today — from code hosting moving inside Cursor, to all three launches getting agents off a single laptop. Thanks so much for listening.

Milo: Been a pleasure. Take care out there, and we'll catch you on the next one.